Collaboration idea: LLM-Wiki as long-term agent-memory companion

[ca1773130n]

Hi! I’m building LLM-Wiki, a local-first implementation of Andrej Karpathy’s LLM Wiki pattern focused on long-lived, agent-native project memory.

First: I really like Understand Anything, and I do not want to fold it into LLM-Wiki or make LLM-Wiki depend on it. It works well as an independent codebase-understanding / graph-generation project.

I think there is a nice loose collaboration surface between the projects:

• Understand Anything: generate a strong code knowledge graph and dashboard for understanding a codebase now.
• LLM-Wiki: compile docs, markdown, research notes, repo context, and local agent sessions into a persistent wiki/static site with agent-facing exports.

A complementary workflow could look like:

# Understand Anything
/understand
# writes .understand-anything/knowledge-graph.json

# LLM-Wiki
llm_wiki project init \
  --source README.md \
  --source docs \
  --source src \
  --source .understand-anything/knowledge-graph.json

llm_wiki project sessions discover --import
llm_wiki project compile
llm_wiki project build-site

The idea is not integration-by-dependency. It is interoperability:

1. Document the knowledge-graph.json shape as a stable exchange artifact.
2. Add or document a schemaVersion if that fits your roadmap.
3. Optionally publish a JSON Schema for downstream tools.
4. Add companion-tool docs/README mentions if both projects agree.
5. Maybe later create an optional external bridge script/repo, outside both cores.

From the current skill docs, the graph shape already looks very usable for this:

project: { name, description, languages, frameworks, analyzedAt, gitCommitHash }
nodes[]: { id, type, name, filePath, summary, tags[], complexity, languageNotes? }
edges[]: { source, target, type, direction, weight }
layers[]: { id, name, description, nodeIds[] }
tour[]: { order, title, description, nodeIds[] }

LLM-Wiki’s side can preserve that graph as part of longer-lived project memory and combine it with:

• markdown/wiki document compilation;
• generated static wiki pages;
• 2D/3D graph website views;
• Claude Code / Codex session history import;
• llms.txt, llms-full.txt, search-index.json, graph.json, per-page .txt/.json siblings;
• optional MCP access for agents.

Would you be open to this kind of lightweight collaboration? I’m happy to start with a small downstream-facing doc PR here, or keep the first step on the LLM-Wiki side and simply document Understand Anything as a recommended companion graph generator.

[Lum1104]

The Understand-Anything also support Karpathy type wiki, can you try /understand-knowledge?